All LLM Models
Browse 47 LLM models with VRAM requirements, quantization options, and hardware compatibility.
Understanding LLM VRAM Requirements
How much VRAM you need depends on the model size and quantization level. Quantization reduces the precision of model weights, trading small quality losses for significantly lower VRAM usage. For example, a 7B parameter model needs ~14 GB at FP16 but only ~4 GB at Q4_K_M quantization.
Model List
DeepSeek V4 Flash
DeepSeek · 290.9B · runs from 124.0 GB
1.5M
2.3KDeepSeek-V4-Flash is the smaller sibling in DeepSeek's V4 series of Mixture-of-Experts language models, with roughly 291 billion total parameters and 13 billion activated per token. Like DeepSeek-V4-Pro, it was pre-trained on more than 32 trillion tokens and post-trained with supervised fine-tuning, reinforcement learning, and distillation, and it supports the same three switchable reasoning-effort modes for trading speed against accuracy on agentic and reasoning tasks. It natively handles a one-million-token context window and is released under the MIT license. Even at 4-bit quantization it needs roughly 167GB of memory, so local use requires multi-GPU or server-class hardware; most users will access it through a hosted endpoint.
DeepSeek R1 0528 Qwen3 8B
DeepSeek · 8.2B · runs from 2.9 GB
With about 8 billion parameters, DeepSeek R1 0528 Qwen3 8B is a dated May 2025 release in DeepSeek's R1 line, built on a Qwen3-based architecture rather than DeepSeek's own base model. It supports explicit step-by-step reasoning, generating intermediate reasoning traces before producing a final answer, which sets it apart from standard instruction-tuned chat models that answer directly. Its modest size means it runs comfortably on modest consumer GPUs, making it accessible for local experimentation with reasoning-style outputs without specialized hardware. It offers a 128K token context window, ample for most reasoning and coding tasks, and is released under the MIT license, allowing broad local and commercial use.
DeepSeek V4 Pro 0813
DeepSeek · 1650.5B · runs from 639.9 GB
DeepSeek-V4-Pro-0813 is DeepSeek's official release of DeepSeek-V4-Pro, superseding an earlier preview build with stronger agentic capabilities aimed at production use. It is a large Mixture-of-Experts model at roughly 1.65 trillion total parameters, built on the V4-Pro architecture with a DSpark speculative-decoding module attached, and it supports configurable reasoning effort (low, high, or max) to control how much it deliberates before answering. It ships with a roughly 1 million token context window and is released under the MIT license, so it can be used and modified freely. At this scale, running it locally means multi-GPU server hardware with hundreds of gigabytes of memory even at 4-bit; most users will access it through a hosted endpoint instead.
DeepSeek R1 Distill Llama 70B
DeepSeek · 70.6B · runs from 20.4 GB
DeepSeek R1 Distill Llama 70B is the largest model in the R1 distillation lineup, combining the reasoning capabilities developed in the full 684.5B R1 with the robust Llama 3.1 70B architecture. At 70 billion parameters, it delivers the strongest reasoning performance of any dense R1 distill, approaching the full R1's quality on many math and coding benchmarks. Running this model locally requires a multi-GPU setup or a single GPU with very high VRAM capacity, though quantized versions can fit on hardware with 48 GB or more. For users who need top-tier open-weight reasoning and have the hardware to support a 70B dense model, this is one of the strongest options available.
DeepSeek v3 0324
DeepSeek · 684.5B · runs from 294.8 GB
DeepSeek V3 0324 is DeepSeek's flagship general-purpose chat model, featuring a 684.5 billion parameter mixture-of-experts architecture with roughly 37 billion parameters active per token. It delivers strong performance across a wide range of tasks including conversation, writing, analysis, coding, and instruction following, competing with the best closed-source models available. Like other large MoE models, V3 requires substantial memory to load all expert weights even though only a fraction are used during inference. Quantized versions make it feasible on multi-GPU setups, and its combination of broad capability with open weights has made it one of the most widely deployed open models for local and self-hosted use.
DeepSeek V4 Flash 0731
DeepSeek · 304.2B · runs from 129.6 GB
DeepSeek V4 Flash 0731 is a large-scale model from DeepSeek's V4 line, weighing in at roughly 304 billion parameters. At this scale it needs multi-GPU or server-class hardware to run; most users will reach it through a hosted endpoint rather than loading it locally on consumer equipment. It is tuned for general chat and instruction-following use. Its standout feature is a 1M token context window, among the largest available in an open-weight model, enabling it to process entire codebases, books, or long conversation histories in a single pass. Released under the permissive MIT license on July 31, 2026, DeepSeek V4 Flash 0731 is aimed at deployments where raw context length and model capacity matter more than local accessibility.
DeepSeek R1 Distill Qwen 1.5B
DeepSeek · 1.8B · runs from 0.8 GB
DeepSeek R1 Distill Qwen 1.5B is the smallest model in the R1 distillation family, packing chain-of-thought reasoning capabilities into just 1.5 billion parameters using the Qwen 2.5 architecture. It represents an ambitious attempt to bring structured reasoning to the smallest practical model size. At this scale, the model can run on virtually any modern GPU and even on CPU-only setups with acceptable speed. While its reasoning depth is naturally limited compared to its larger siblings, it still demonstrates structured thinking patterns that set it apart from generic models of similar size.
DeepSeek R1 Distill Qwen 7B
DeepSeek · 7.6B · runs from 3.0 GB
DeepSeek R1 Distill Qwen 7B compresses the reasoning techniques from DeepSeek's full R1 model into a compact 7.6 billion parameter dense model built on the Qwen 2.5 architecture. Despite its small footprint, it demonstrates surprisingly capable step-by-step reasoning on math and logic problems that would stump many models several times its size. This is one of the most accessible reasoning models available for local use, fitting comfortably on GPUs with 6 GB or more of VRAM when quantized. It strikes a practical balance between genuine chain-of-thought reasoning ability and the hardware constraints of a typical consumer setup.
DeepSeek R1 Distill Qwen 32B
DeepSeek · 32.8B · runs from 9.8 GB
DeepSeek R1 Distill Qwen 32B takes the reasoning capabilities developed in the full 684.5B R1 model and distills them into the 32.8 billion parameter Qwen 2.5 architecture. The result is a dense model that punches well above its weight class on math, science, and coding reasoning tasks, often matching models two to three times its size. At around 32.8 billion parameters, this model fits comfortably on a single high-end consumer GPU when quantized to 4-bit precision, making it one of the most capable reasoning models you can run on a desktop workstation.
Deepseek Coder 6.7B Instruct
DeepSeek · 6.7B · runs from 4.2 GB
DeepSeek Coder 6.7B Instruct is a first-generation code-specialized model trained on a large corpus of source code and programming-related data. At 6.7 billion parameters, it provides solid code completion, generation, and explanation capabilities across popular programming languages while remaining small enough to run on most consumer GPUs. While newer models in the DeepSeek lineup have surpassed it in raw capability, this model remains a practical choice for users who need a lightweight local coding assistant with minimal hardware requirements. It runs well on GPUs with as little as 6 GB of VRAM when quantized.
DeepSeek R1 0528
DeepSeek · 684.5B · runs from 192.1 GB
DeepSeek R1 0528 is an updated release of the R1 reasoning model, incorporating improvements to training and inference that sharpen its performance on complex multi-step problems. It retains the same 684.5 billion parameter mixture-of-experts architecture as the original R1, with approximately 37 billion parameters active per forward pass. This revision addresses several edge cases where the original R1 struggled, delivering more consistent reasoning chains and fewer hallucinations on difficult math and coding tasks. Hardware requirements remain identical to the original R1, so users already set up to run the first version can swap in the 0528 weights with no changes to their infrastructure.
DeepSeek R1 Distill Qwen 14B
DeepSeek · 14.8B · runs from 5.1 GB
DeepSeek R1 Distill Qwen 14B sits in a sweet spot between the smaller 7B distill and the more demanding 32B version, offering strong reasoning performance at 14.8 billion parameters on the Qwen 2.5 architecture. It captures a meaningful share of the full R1's chain-of-thought capabilities while keeping resource requirements within the range of mainstream consumer GPUs. Quantized to 4-bit, it fits comfortably on GPUs with 12 GB of VRAM, delivering reliable step-by-step reasoning for math, logic, and analytical problems.
DeepSeek V4 Flash Vision Exp
DeepSeek · 304.6B · runs from 84.1 GB
DeepSeek V4 Flash Vision Exp is an experimental multimodal model from DeepSeek, a mixture-of-experts design with roughly 304.6 billion total parameters, only a small fraction of which are active per token. It extends the text-only DeepSeek V4 Flash line with image understanding, built for agentic tasks, reasoning, and tool use. Only the active parameters are computed per token, keeping decoding fast, but all weights must fit in memory, making this a server-class model best run across multiple GPUs or a hosted endpoint. The model supports a 1,048,576 token context window, suited to very long documents or codebases. It is released under the MIT license, one of the most permissive open licenses, with no restrictions on commercial use. Published in August 2026 as an experimental release, it brings image input to the Flash line.
DeepSeek R1 Distill Llama 8B
DeepSeek · 8.0B · runs from 2.8 GB
DeepSeek R1 Distill Llama 8B brings R1's reinforcement-learned reasoning capabilities to the widely supported Llama 3.1 8B architecture. By distilling the full 684.5B R1 model's reasoning patterns into this 8 billion parameter dense model, DeepSeek created a version that benefits from the extensive Llama ecosystem of tools, quantizations, and inference engines. For users who prefer the Llama architecture or already have tooling built around it, this model offers a plug-and-play path to chain-of-thought reasoning. Its hardware requirements are very approachable, running well on consumer GPUs with 8 GB or more of VRAM at common quantization levels.
DeepSeek V4 Pro
DeepSeek · 1598.8B · runs from 679.8 GB
DeepSeek-V4-Pro is a preview release in DeepSeek's V4 series of Mixture-of-Experts language models, built for highly efficient million-token-context reasoning. It totals roughly 1.6 trillion parameters with 49 billion activated per token, pre-trained on more than 32 trillion tokens and post-trained through supervised fine-tuning, reinforcement learning, and on-policy distillation. This instruct release supports three switchable reasoning-effort modes, from fast everyday answers up to an extended "Think Max" mode for harder problems, plus stronger tool-calling and agentic behavior. It natively handles a one-million-token context window and is released under the MIT license. At this scale, running it locally means server-class or multi-GPU hardware, with 4-bit quantization still needing roughly 900GB of memory; most people will reach it through a hosted endpoint instead.
DeepSeek V4.1 Flash
DeepSeek · 763.2B · runs from 324.7 GB
DeepSeek-V4.1-Flash is DeepSeek's multimodal Mixture-of-Experts model, processing images and text together and generating text autoregressively. It uses a Causal Encoder-Decoder architecture with roughly 763 billion total parameters, but activates only about 8 billion during prefill and 16 billion during decode, keeping input-heavy agentic workloads efficient. The model supports a continuously adjustable reasoning-effort setting that trades inference cost for accuracy, and was trained with aggressive KV-cache compression aimed at long-context, tool-using agents. Context reaches up to one million tokens, and it is released under the MIT license. Given its size, local use calls for multi-GPU or server-class hardware, with roughly 439GB of memory needed even at 4-bit; most users will rely on a hosted API instead.
DeepSeek R1
DeepSeek · 684.5B · runs from 192.1 GB
DeepSeek R1 is a groundbreaking reasoning model that uses reinforcement learning to develop chain-of-thought capabilities without relying on supervised fine-tuning. With 684.5 billion total parameters in a mixture-of-experts architecture (only 37 billion active per token), R1 achieves performance competitive with OpenAI's o1 on math, coding, and complex reasoning benchmarks while remaining fully open-weight. Running the full R1 locally is a serious undertaking, requiring well over 300 GB of VRAM at full precision, though quantized versions bring it within reach of multi-GPU setups. For users who want R1-level reasoning on more modest hardware, DeepSeek also released a family of distilled models that pack R1's reasoning patterns into smaller dense architectures.
Deepseek Coder 6.7B Base
DeepSeek · 6.7B · runs from 3.2 GB
Deepseek Coder 6.7B Base is a 6.7B-parameter open language model from DeepSeek in the DeepSeek Coder family. It supports a context window of up to 16,384 tokens. See its VRAM requirements by quantization and which GPUs and Macs can run it locally below.
Deepseek Coder 33B Instruct
DeepSeek · 33.3B · runs from 14.6 GB
DeepSeek-Coder-33B-Instruct is DeepSeek's 33.3-billion-parameter instruction-tuned code model, initialized from DeepSeek-Coder-33B-Base and further fine-tuned on 2 billion tokens of instruction data for chat-style code generation, completion, and fixing. The underlying Deepseek Coder family was pretrained from scratch on 2 trillion tokens (87% code, 13% natural language in English and Chinese) with project-level context and a fill-in-the-blank training objective, giving it strong project-level completion and infilling ability alongside state-of-the-art open-model results on HumanEval, MBPP, and related coding benchmarks at the time of release. At 33.3 billion parameters, it needs a high-end consumer GPU or a multi-GPU setup once quantized. Context length is 16,384 tokens. Model weights are released under DeepSeek's custom model license, which permits commercial use. It was published in November 2023, alongside 1.3B, 5.7B, and 6.7B siblings and the corresponding base checkpoint.
DeepSeek Coder v2 Lite Instruct
DeepSeek · 15.7B · runs from 7.2 GB
DeepSeek Coder V2 Lite Instruct is a code-focused mixture-of-experts model with 15.7 billion total parameters, trained to handle both programming tasks and general conversation. It supports a wide range of programming languages and excels at code generation, debugging, explanation, and refactoring. The MoE architecture keeps compute costs manageable despite the model's broad capabilities, and the Lite variant is sized to run on a single consumer GPU. For developers looking for a capable local coding assistant that can also handle general chat, this model offers an appealing combination of code specialization and practical hardware requirements.
DeepSeek V3.1
DeepSeek · 684.5B · runs from 192.1 GB
DeepSeek-V3.1 is a hybrid instruct model from DeepSeek that supports both thinking and non-thinking modes within a single set of weights, switched via its chat template. It is a roughly 684-billion-parameter Mixture-of-Experts model with 37 billion parameters activated per token, post-trained on top of DeepSeek-V3.1-Base with an extended long-context training phase, adding improved tool calling and dedicated formats for code and search agents. It has a 163,840-token context window and is released under the MIT license. Even at 4-bit quantization it needs roughly 394GB of memory, so local deployment requires server-class or multi-GPU hardware rather than a single machine.
DeepSeek OCR 2
DeepSeek · 3.4B · runs from 1.9 GB
DeepSeek OCR 2 is DeepSeek's compact vision-language model built for optical character recognition and document understanding, totaling 3.4 billion parameters with about 1.2 billion active per token through its mixture-of-experts design. Only the active parameters compute per token, keeping inference fast, though the full weight set still needs to fit in memory; at this size that's within reach of a consumer GPU or laptop once quantized. It uses a DeepEncoder V2 vision architecture that reasons over document layout semantically instead of scanning images in a fixed pattern. The model has an 8K token context window, suited to single-document OCR passes. It is released under the Apache 2.0 license, allowing unrestricted commercial and research use, and was published in January 2026 as DeepSeek's OCR successor.
DeepSeek V2.5
DeepSeek · 235.7B · runs from 67.7 GB
DeepSeek-V2.5 is DeepSeek's merged general-purpose and coding chat model, combining DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct into a single mixture-of-experts checkpoint with about 235.7 billion total parameters and roughly 21.4 billion active per token. DeepSeek reports it better aligns with human preferences and improves writing and instruction-following over its two predecessors, with gains on AlpacaEval, ArenaHard, and coding benchmarks like HumanEval and LiveCodeBench. It is a chat, not base, model. Given its total parameter count, DeepSeek states that running it in BF16 requires eight 80GB-class GPUs, so it needs a substantial multi-GPU server, not a single consumer machine, even once quantized. Context length is 163,840 tokens. Model weights are released under DeepSeek's custom model license, which permits commercial use, while the surrounding code is MIT-licensed. It was published in September 2024, unifying the separate V2-Chat and Coder-V2-Instruct lines into one model.
Deepseek Coder 33B Base
DeepSeek · 33.3B · runs from 14.6 GB
DeepSeek-Coder-33B-Base is DeepSeek's 33.3-billion-parameter base (pretrained, not instruction-tuned) code model, trained from scratch on 2 trillion tokens (87% code, 13% natural language in English and Chinese) using grouped-query attention, project-level code context, and a fill-in-the-blank training objective for project-level completion and infilling. It is intended as a foundation for further fine-tuning or direct code completion and infilling use, not conversational use; the instruction-tuned DeepSeek-Coder-33B-Instruct is built on top of it. At release, the Deepseek Coder family achieved state-of-the-art open-model results on HumanEval, MultiPL-E, MBPP, DS-1000, and APPS. At 33.3 billion parameters, it needs a high-end consumer GPU or a multi-GPU setup once quantized. Context length is 16,384 tokens. Model weights are released under DeepSeek's custom model license, which permits commercial use. It was published in October 2023, alongside 1.3B, 5.7B, and 6.7B siblings and the corresponding instruct checkpoint.
DeepSeek v2 Lite Chat
DeepSeek · 15.7B · runs from 5.1 GB
DeepSeek v2 Lite Chat is a 15.7B-parameter open language model from DeepSeek in the DeepSeek V2 family. It supports a context window of up to 163,840 tokens. See its VRAM requirements by quantization and which GPUs and Macs can run it locally below.
DeepSeek V3.1 Terminus
DeepSeek · 684.5B · runs from 192.1 GB
DeepSeek-V3.1-Terminus is an updated checkpoint of DeepSeek-V3.1, a Mixture-of-Experts chat and reasoning model with roughly 40.1 billion active parameters out of about 684.5 billion total, fine-tuned from DeepSeek-V3.1-Base. This revision keeps the same capabilities as V3.1 while fixing issues reported by users, chiefly reducing mixed Chinese-English text and stray characters in output and further improving the model's Code Agent and Search Agent performance, including gains on BrowseComp, SWE-bench Verified, and Terminal-bench. Its scale requires a multi-GPU server; it is not something that fits on consumer hardware. Context length is 163,840 tokens. It is released under the MIT license, permitting unrestricted commercial and research use. It was published in September 2025, as a refinement of DeepSeek-V3.1 rather than a new base model.
DeepSeek V3.2
DeepSeek · 685.4B · runs from 192.3 GB
DeepSeek V3.2 is the latest iteration of DeepSeek's general-purpose flagship, building on the V3 architecture with 685.4 billion total parameters in a mixture-of-experts configuration. This update refines the model's conversational abilities, instruction following, and multilingual performance compared to earlier V3 releases. Running V3.2 locally requires significant GPU resources due to the large total parameter count, though the MoE design means only a subset of parameters are active for any given token. Users with multi-GPU workstations or servers can run quantized versions effectively, making this one of the most powerful open-weight chat models available for self-hosted deployment.
DeepSeek V2.5 1210
DeepSeek · 235.7B · runs from 67.7 GB
DeepSeek-V2.5-1210 is an updated checkpoint of DeepSeek's DeepSeek-V2.5 chat model, a mixture-of-experts model that routes across 160 experts (plus 2 always-on shared experts) with 6 selected per token, for roughly 21.4 billion active parameters out of about 236 billion total. Compared to the original DeepSeek-V2.5, this revision improves math (MATH-500 from 74.8% to 82.8%), coding (LiveCodeBench from 29.2% to 34.38%), and writing and reasoning quality, plus smoother file-upload and webpage-summarization behavior. The card states that full-precision inference needs an eight-GPU server, so local use realistically means a multi-GPU setup even after heavy quantization. Context length is 163,840 tokens. It is released under DeepSeek's custom Model License Agreement, and was published in December 2024, about a year before the DeepSeek-V3 family.
DeepSeek Coder v2 Instruct
DeepSeek · 235.7B · runs from 73.5 GB
DeepSeek-Coder-V2-Instruct is DeepSeek's mixture-of-experts code and math model, continuing pretraining from an intermediate DeepSeek-V2 checkpoint with an additional 6 trillion tokens focused on code. It has about 235.7 billion total parameters with roughly 21.4 billion active per token, and DeepSeek reports coding and mathematical reasoning performance comparable to GPT-4 Turbo, expanding supported programming languages from 86 to 338 versus the original DeepSeek-Coder line. It is an instruction-tuned chat model, not a base model, intended for code generation, completion, and reasoning tasks. Given its total size, DeepSeek recommends an eight-GPU, 80GB-class server for BF16 inference, so it needs a substantial multi-GPU setup even once quantized. Context length is 163,840 tokens; DeepSeek's own materials describe this as an extension "from 16K to 128K" over the original Deepseek Coder. Model weights are released under DeepSeek's custom model license, which permits commercial use, while the surrounding code is MIT-licensed. It was published in June 2024.
DeepSeek v3
DeepSeek · 684.5B · runs from 294.8 GB
Released in December 2024, DeepSeek V3 is a large chat-oriented language model from DeepSeek with roughly 684 billion parameters, placing it among the largest openly released checkpoints of its generation. It is built for general conversational and instruction-following use rather than a narrow specialty. With a 160K token context window, DeepSeek V3 can process long documents and extended multi-turn conversations in a single pass without truncation. At this parameter scale, it needs multi-GPU or server-class hardware to run at all; most users will reach it through a hosted endpoint rather than attempting to run it locally on their own machine.